Generative AI has made it possible to automate almost anything. That’s exactly why you need a better filter for what you should tackle with customer service automation.

Last month, something in my HelloFresh order was wrong. I opened the app, reported the issue, and in under a minute, I was a happy customer again.

Tactically, the solution was simple, offering me a modest account credit, but the experience was positive because of three things:

1. Effectiveness: They made clear the credit wasn’t a take-it-or-leave-it offer—if it felt too small, I could push back.
2. Ease: They credited my account on the spot.
3. Emotion: They believed me, no photos or explanations required.

The thing is, I’ve never once pushed back. Just knowing I could is enough to make me feel taken care of.

This thirty-second interaction is a masterclass in automation done well. It was easy and effective, and the trust and flexibility created positive emotions. It’s a stark contrast to most automations, which fall into the trap of prioritizing “efficiency” over all else. In doing so, they may resolve the transactions, but they quietly erode the relationship.

---

---

Efficiency and consistency are great goals for the boardroom. And generally terrible goals for customers. The moment a customer’s situation doesn’t match the script—which is precisely when they need the brand most—rigid automations stop saving money and start costing serious brand equity, putting future revenue at risk.

Generative AI raises the stakes on this dramatically. Why? Consider the rubric below:

This rubric represents traditional wisdom. It’s not wrong, but AI is the first technology that’s come anywhere close to showing it can start to do more things on the right. Now, are those appearances more than skin deep? For now, mostly not. But the writing is on the wall: More—and potentially better—automation is coming.

That shifts the question for business leaders from “can we automate this?” at a functional level to “(how) should we automate this?” at an experience level.

The Low-Hanging Fruit You’re Still Ignoring

Before reaching for generative AI, start with the unsexy wins: RPA, cleaner workflows, better-designed forms. There’s real value sitting untouched here. These are the interstitial or behind-the-scenes steps where customers are still doing work your systems should be doing for them.

If you’re not confident you’ve captured all the easy wins, here are two good places to look:

• Account Information. Eliminate duplicate asks. If a customer has told you something once, “write once, update everywhere” should mean they never have to tell you again.
• Proactive Status and Reminders. Uncertainty is its own form of friction. Such timely, unprompted updates cost little and buy real goodwill.

None of this requires AI. All it requires is admitting these moments matter more than their size suggests.

Efficiency Is the Wrong First Question

Once you move past the obvious fixes, the real design question isn’t “how do we automate this faster?” but “what are we actually optimizing for?” Get the metric wrong, and every automation you build afterward goes in the wrong direction.

Consider a credit card company that measures customer service success by average handle time. Shorter calls look better on the dashboard. But if the business’s actual goal is increasing card usage, a slightly longer call—one where a rep takes an extra ninety seconds to explain a benefit the customer didn’t know they had—can generate more long-term value than a dozen calls optimized purely for speed.

Two patterns show up again and again when businesses optimize for the wrong thing:

• Closures Over Resolutions. Any incentive structure that rewards marking something “done” faster than it rewards actually solving it.
• Wins Over Worth. Counting a sale or a lead without weighing what that customer is potentially worth over the following two years.

Nearly one-third of consumers have stopped buying from a brand because of poor customer experience, and even more have done so because of negative product or service experiences.

Being honest about this usually surfaces a fair number of penny-wise, pound-foolish decisions already baked into how automation gets prioritized. Take a second look before you leap.

What “Good” Actually Means to Customers

The business side of this equation is measurable. The customer side is less obvious, but there’s a clear pattern. Forrester’s research on CX quality shows three things that matter, and all three have to be present:

• It was effective—it solved the actual problem.
• It was easy—it didn’t cost the customer more effort than they expected.
• It created a positive emotional response—it met expectations of the brand and improved the customer’s situation.

That emotional response is what automation strategies chronically miss, mostly because it doesn’t fit on a dashboard.

Here’s the reassuring part: your customers are not infinitely unique or unpredictable. Map how people actually behave across your key use cases, and clear patterns will emerge. One useful way to find them is to work backwards instead of forwards: instead of starting from how a process is designed and asking where to bolt on automation, start from where things are actually failing today and ask why. Some of what you find will be automatable. Some won’t be—and that’s okay. Both answers are useful because, either way, you now know where to route the experience with intention instead of by default.

Once you’ve found candidates for automation, three questions determine whether (and how) you should actually automate them.

Question 1: Does This Require Judgment Outside Policy?

Most organizations discover the same thing when they try to automate seriously: how much undocumented knowledge and invisible human judgment was actually holding the process together. Policy is rarely as airtight as leadership assumes, and even where it is, training and enforcement have gaps. The result is plenty of processes that look automatable on paper but stay stubbornly manual in practice because the real rules were never written down.

The straightforward fix is to go watch how the work actually gets done, find what isn’t captured in policy, and codify it.

The more valuable fix is harder: change the policy itself. A surprising number of process steps exist purely out of legacy habit. Before asking “how do we automate this?” it’s worth asking “do we need this at all?“

Keep discretion human when the outcome won’t change but the customer will hate it. What matters isn’t the answer, but whether the customer feels respected, understood, and confident in how they got it. For your highest-value customers in these moments, simply being heard by a person can preserve (or even improve) how they feel about the brand, and skilled handling of a hard moment often opens the door to an upsell or cross-sell opportunity that a purely automated flow could never create.

Question 2: Is This a Peak or an Ending?

People remember two moments: the emotional peak and the end. Journey mapping finds these, but most companies still underinvest, especially in segmenting by what customers actually need emotionally, rather than treating everyone the same.

You don’t need a full mapping exercise to get started. A look at your top call drivers, your complaint patterns, or a two-hour workshop with people who talk to customers daily will surface a solid shortlist of emotionally loaded moments fairly quickly. From there, look for ways to identify those moments earlier and route them toward the kind of support that can actually deliver on effectiveness, ease, and emotion. Sometimes it’s something as simple as one well-placed question in an IVR.

While peak moments are highly contextual, endings are an easier fix, mostly because they go unmanaged by default. For example, a request to cancel a subscription after a death in the family gets the identical “we’re sad to see you go” email sent to someone who just got bored—or worse, a “please come back” discount to save the unsavable.

You almost certainly have the data to see these moments coming. The opportunity is building the automation that actually meets them with a response that respects the customer’s emotional context instead of a tone-deaf procedural message.

Question 3: What Happens If You Get It Wrong?

Every business accepts some level of risk, whether it’s shrinkage, chargebacks, or write-offs. But automation makes us expect perfection. The better question to ask is, “what level of risk is acceptable when something goes wrong?” Three design questions get you there:

1. How fast is detection? This matters more than whether errors happen at all. An error caught in an hour is a minor operational cost. The same error caught a month later, after it’s compounded across ten thousand customers, is a headline.
2. How big is the blast radius? Define the boundaries of what an automated decision is allowed to get wrong before you deploy it. A refund automation capped at fifty dollars fails safely. Worst case, you’ve given away fifty dollars to someone who didn’t earn it.
3. How cheap is the undo? Every automated decision should ship with a reversal built in, and the more self-service it is, the better. Credit card issuers already do this with provisional credit on disputed charges: the system acts immediately and unwinds later if the dispute turns out to be unfounded.

HelloFresh’s refund flow is, again, the model—a process engineered to be cheap to run, cheap to reverse, and fast to catch if it’s being abused.

The best safety net for automation is, you guessed it, skilled humans. Nearly 80% of consumers strongly prefer interacting with a human rather than an AI agent for customer service, and 89% believe companies should always offer the option to speak with a person. This is why, as you ramp up automation, you need to equally ramp up the quality and empowerment of your human support to handle the exceptions. And routing those exceptions should happen faster than ever through automations that fail gracefully.

Automation Is a Tool. Trust Is the Outcome.

Generative AI makes it tempting to ask how much you can automate. That’s the wrong question. The right one is how to automate in ways that demonstrate and create shared value between your business and your customers.

Moving past efficiency means weighing effectiveness, ease, and—most importantly—emotion together, and it means being deliberate about where automation belongs and where it doesn’t.

Practically, that starts in three places:

1. Policies worth changing or codifying.
2. Sensitive moments worth routing to a human.
3. A system for catching and undoing mistakes cheaply enough that “acceptable risk” replaces “expected perfection” as the standard you’re building toward.

Do that, and automation stops being something to cut costs and becomes something to create value—which, it turns out, is the only version of it that actually pays off.

The future won’t belong to the companies that automate the most. It will belong to the companies that know where human judgment, empathy, and trust create more value than efficiency alone. That’s the reality—or new reality, if you will—business leaders need to navigate as AI transforms every corner of the enterprise.